Does xG Predict Results Better Than EPV?
Using EPV and xG to separate football performance from match outcome noise.
The following summary critically reviews the research paper titled “AI in Bundesliga match analysis—expected possession value (EPV) vs. expected goals (xG) to predict match outcomes in soccer” by Leander Forcher, Leon Forcher, Alexander Woll, and Stefan Altmann. All data, figures, and analysis presented here are drawn from their original work; I do not claim any authorship or ownership of the content. This summary has been written to provide a concise and technically informed synthesis of the paper’s findings, methodologies, and implications, while maintaining fidelity to the authors’ intellectual contributions.
1 Introduction
The paper begins from a central tension in football analytics: football is highly structured, but match outcomes remain noisy. The authors frame this through the idea that “it is still chance that dominates the results in the game of soccer” [1], using Arminia Bielefeld’s 2025 DFB-Pokal run as an example of how weaker teams can still progress against stronger opponents. This uncertainty is not presented as a weakness of analysis, but as the reason why clubs seek better models. The role of data science is to help analysts and coaching staff “tip the probabilities in their team’s favor” by identifying signals that contain more stable information than the final score alone.
Within this context, the paper positions expected goals (xG) as one of the most established KPIs in football analytics. xG is useful because it evaluates the quality of shots, but it is also limited because it only sees the game if a shot has happened. The authors therefore contrast xG with expected possession value (EPV), which attempts to value every possession or game state by estimating the probability of scoring from that situation.
The broader problem is not simply whether xG or EPV is “better”, but how to judge the value of any KPI in a landscape where football analytics now produces a large number of metrics. The authors argue that “it is crucial to identify which KPIs contain the most valuable information” [10], and define predictive power as one meaningful way to assess this value. If a KPI helps predict match outcomes or future performance, then it likely captures relevant information about team strength or match performance.
The introduction also separates two different prediction settings.
Pre-match prediction only uses information available before the match, making it relevant for opponent analysis, forecasting, and trend assessment.
Post-match prediction uses information from the match itself, making it more relevant for evaluating what happened and separating performance from randomness.
The paper’s stated contribution is therefore to compare xG and EPV in both settings, asking whether shot-based information or possession-based information better predicts match outcomes when used before and after the game.
2 Methods
The methodology is built around four match outcome prediction approaches: xG pre-match, EPV pre-match, xG post-match, and EPV post-match. This design lets the authors compare not only the two metrics, but also the analytical context in which each metric is used. Pre-match models ask whether recent xG or EPV information helps forecast the next match. Post-match models ask whether the xG or EPV produced inside a match better explains the result.
The study follows an applied machine learning workflow. The authors use official Bundesliga event and tracking data, compute xG, EPV, and xBG information, build pre-match goal prediction models, convert those predicted goals into match outcome probabilities, and evaluate the quality of the resulting forecasts using probabilistic scoring rules. Their approach is therefore not a pure betting model and not a descriptive match report model. It sits between forecasting and performance analysis.
2.1 Data
The dataset consists of official event and tracking data from 918 Bundesliga matches across three seasons: 2022/23, 2023/24, and 2024/25. The event data came from Sportec Solutions and included around 30 event types with more than 100 attributes, following the German Football League’s official match data catalogue [22]. The tracking data was collected at 25 Hz using TRACAB and included both ball and player positions.
The authors emphasize that the event and tracking data were synchronized, which is important because the models depend on combining discrete actions with continuous spatial context. This is especially relevant for EPV and xBG, where the value of a situation depends not just on the event being performed, but on player locations, pressure, passing options, defensive organization, and the broader match state.
2.2 xG/EPV/xBG
The paper relies on three underlying machine learning models: expected goals, expected possession value, and expected ball gain. xG and EPV represent attacking value, while xBG provides a defensive counterpart by estimating the probability that the defending team will regain possession.
This combination is important because the EPV-based models do not only ask how likely a team is to score. They can also include information about the risk of losing the ball. In that sense, the paper treats football possessions as risk-reward situations: an attack may carry high scoring value, but it may also expose the team to a high probability of turnover.
2.2.1 xG
The xG model estimates the probability that a shot becomes a goal. The authors refer to existing reviews of xG models by Cavus and Biecek [25] and Hewitt and Karakus [4], while using an existing model from Anzer and Bauer [24]. In the paper’s framing, xG is a strong post-match measure because shots are the direct route through which almost all goals are scored.
The key limitation is also clear. Because xG only activates once a shot occurs, it ignores valuable attacking sequences that do not end with a shot. A team may create promising entries, overloads, or dangerous possession chains, but if the final pass fails or the shot never comes, a pure xG model sees little or nothing.
2.2.2 EPV
EPV estimates the probability of scoring “in the following seconds of every match situation on the pitch” [8, 26]. In this study, EPV is defined as the probability of scoring within the next ten seconds from a given game state. This gives it a broader analytical scope than xG, because it can value entire possessions and not only the final shot.
The EPV model is trained separately for different playing phases, including offensive play, offensive transitions, direct free kicks, indirect free kicks, corners, and throw-ins. This is a useful methodological choice because the value structure of a corner is not the same as the value structure of open play or transition. The authors note that the model uses 38 features capturing match situation, attacking performance, and defensive performance. Examples include distance to goal, relative pitch position, passing options, deep runs, defensive pressure, and team organization.
This makes EPV closer to a full attacking process model. It can detect danger before the shot, and it can retain information about attacks that xG would miss. That becomes especially important in the pre-match setting, where recent attacking process may be more predictive than recent shot quality alone.
2.2.3 xBG
xBG complements EPV by estimating the probability that the defending team will regain possession in every match situation. The authors use xBG to represent defensive risk, meaning that a possession can be evaluated not only by its scoring potential but also by the likelihood of losing the ball.
This creates a richer interpretation of possession value. The attacking team’s reward is captured by EPV, while the attacking team’s risk is captured through the opponent’s xBG. For match analysis, this is important because not all dangerous-looking attacks are equally efficient or safe. Some carry high value but also high turnover risk, while others may create moderate value with better control.
2.3 Pre-match predictions
The pre-match approach uses a two-stage framework. First, machine learning models predict the number of goals each team is expected to score in the upcoming match. Second, those predicted goal numbers are converted into outcome probabilities through match simulations.
The models only use information available before the match. This includes recent team performance from the last three matches, contextual information such as venue, opponent strength, attacking performance, defensive performance, and the difficulty of recent opponents. This makes the pre-match setting a test of whether xG or EPV carries more predictive information about future matches.
2.3.1 Features
The feature design follows the logic introduced by Dixon and Coles [15], where a good match prediction model should account for team ability, home advantage, recent performance, attacking and defensive strength, and the quality of recent opponents. The paper’s feature set reflects this structure.
The authors include match context through venue, opponent strength through table position and points differences, and recent match difficulty through the strength of the last three opponents. They also include attacking performance for the considered team and defensive performance for the upcoming opponent. In the EPV-based version, this includes EPV and xBG variables; in the xG-based version, this includes xG variables. Table 2 is important here because it shows that the models are not simply comparing raw xG and raw EPV. They are comparing broader feature families built around each metric.
The feature categories allow the model to distinguish between, for example, a team that has accumulated strong EPV against weak opposition and a team that has done so against stronger opposition. This is essential for football forecasting because recent performance is not equally informative in all contexts.
2.3.2 Modelling of number of goals (xGoalNumber)
The goal prediction task is defined at team-match level. For each match, each team becomes one sample, and the model predicts how many goals that team will score against the upcoming opponent. Because the models need recent performance history, the first six matchdays are excluded. This leaves 504 matches and 1,008 team-level samples.
The authors train Random Forest and XGBoost regression models for both the EPV and xG pre-match approaches. In the EPV pre-match model, xG-related variables are excluded. In the xG pre-match model, EPV- and xBG-related variables are excluded.
The data split is performed match by match, using an 80/20 hold-out split to avoid leakage between samples from the same match. On the training set, the authors apply five-fold cross-validation and randomized hyperparameter tuning, optimizing for RMSE. They then use SHAP values to interpret the selected models. SHAP is used because it can estimate how much each feature contributes to model predictions, both globally and for individual cases.
2.3.3 Prediction of match outcomes
Once the models predict each team’s expected number of goals, the authors use independent Poisson distributions for home and away goals, following Maher [31]. The maximum score is capped at eight goals, based on the observed dataset. For each match, they simulate 10,000 possible outcomes.
This step transforms goal predictions into probabilities of home win, draw, and away win. It also allows the model to produce more detailed scoreline distributions, rather than only predicting the most likely outcome. The practical value is that analysts can inspect not only which team is favoured, but also how the probability mass is distributed across possible goal differences and scorelines.
2.3.4 Baseline approach
The pre-match models are compared against an Elo-based baseline. This baseline uses a Bradley–Terry–Davidson framework to account for draws, with an average Bundesliga draw rate of 0.25 from the 2022/23 and 2023/24 seasons. It also includes a fixed home advantage of 60 Elo points, a k-factor of 20, and all teams start from an Elo rating of 1,500.
The purpose of this baseline is not to create the strongest possible forecasting system, but to provide a simple, interpretable comparison point. Since Elo is widely used and easy to understand, outperforming it gives the authors a clearer benchmark for judging whether the richer xG and EPV models add value.
2.4 Post-match predictions
The post-match approach is simpler and more direct. For xG, each shot is treated as a scoring opportunity with a modelled probability of becoming a goal. For EPV, each possession is treated as an opportunity with a modelled probability of producing a goal. For each match, the authors run 10,000 simulations to generate possible outcomes from the observed performance.
This design makes the post-match models a tool for judging what the match performance “deserved” in probabilistic terms. The xG post-match model asks what results would likely occur if the same shots were repeated many times. The EPV post-match model asks what results would likely occur if the same possession values were repeated many times.
2.5 Evaluation
The authors evaluate the models using Ranked Probability Score, accuracy, and Brier Score. RPS is especially suitable because match outcome prediction is ordinal: home win, draw, and away win are not just three unrelated classes, since draw lies between the two win outcomes in goal-difference terms. Lower RPS values indicate better probabilistic forecasts.
Accuracy is computed by taking the outcome with the highest predicted probability and checking whether it matches the actual result. This is intuitive, but less informative than probabilistic scoring because it ignores whether a model was cautiously or confidently wrong. The Brier Score measures the squared difference between predicted probabilities and actual outcomes, and the authors decompose it into uncertainty, reliability, and resolution.
For the intermediate goal prediction models, they also report MAE, MSE, RMSE, and R^2. These metrics evaluate how well the first-stage model predicts the number of goals before those predictions are transformed into outcome probabilities.
3 Results
The match outcome prediction dataset contains 756 matches, covering matchdays 7 to 34 across the three seasons. Home teams won 335 matches, or 44.3%, while 195 matches ended in draws, or 25.8%, and away teams won 226 matches, or 29.9%. Home teams scored 1.79 ± 1.45 goals per match, while away teams scored 1.36 ± 1.19.
The most common scoreline was 1:1, followed by 2:1 home wins, 2:0 home wins, and 1:2 away wins. These distributions matter because they shape the baseline difficulty of the prediction task, especially around draws.
For pre-match prediction, the EPV-based XGBoost model produced the best performance. It achieved RPS = 0.194 and Accuracy = 0.583. The xG pre-match models were slightly weaker, with the best xG pre-match result achieving RPS = 0.199 and Accuracy = 0.556. The Elo baseline achieved RPS = 0.202 and Accuracy = 0.553, meaning the EPV pre-match model outperformed both the xG pre-match model and the simple Elo benchmark.
For post-match prediction, xG clearly performed best. The xG post-match approach achieved RPS = 0.148 and Accuracy = 0.656, outperforming all other approaches. The EPV post-match approach reached RPS = 0.191 and Accuracy = 0.596. This result suggests that once the match has already been played, the shot-based information captured by xG provides a more direct explanation of the result than possession-level EPV.
The figures reinforce this interpretation. Figure 1 shows feature importance and SHAP values for the EPV and xG pre-match XGBoost models.
Figure 2 compares calibration curves and probability distributions for the EPV pre-match model and the xG post-match model.
Figure 3 later demonstrates how the two approaches can be used together in practical match analysis, comparing pre-match expectations with post-match performance.
4 Discussion
The discussion interprets the main finding as a contextual answer rather than a single winner between xG and EPV. In pre-match analysis, EPV performs better. In post-match analysis, xG performs better. The authors therefore argue that the value of a KPI depends on the question being asked.
The pre-match result suggests that EPV captures a broader and more predictive picture of team performance. Because EPV evaluates every possession and not only shots, it can include information from attacks that did not end in shots but still reflected team quality, territorial control, chance creation processes, and pressure on the opponent. The authors argue that EPV “may therefore provide a more comprehensive representation of a team’s scoring potential of an entire attack even in the absence of a shot and a resulting xG value.”
The post-match result points in the opposite direction. Once a match has been played, xG is more informative because shots are the immediate actions from which goals are normally scored. The authors note that “shots are practically the only way to score (except from own goals).” Therefore, the quality and quantity of shots produced in the actual match provide a strong basis for evaluating the result. xG is reductionist, but in this setting that reduction is useful. It filters the game down to the decisive scoring chances.
The SHAP analysis adds further detail. Opponent strength is one of the most important feature groups in both models. Differences in points and table position strongly affect predicted goals, with points appearing more informative than rank because they provide a more granular signal. Match venue is also highly important, reflecting the well-established home advantage in football [40].
Recent attacking performance also matters, but not all attacking metrics behave the same way. Goal difference over the last three matches is more informative than raw goals, while the number of goals alone shows little predictive power. This supports the idea that goals are noisy and contain more randomness than process-based metrics. In the EPV model, the number of possessions where EPV exceeds 0.5 is especially important, suggesting that repeated creation of very high-value possession states carries meaningful predictive signal. In the xG model, total xG over the last three matches is the most important xG-specific feature, but it ranks lower than the strongest EPV feature.
The authors also highlight the relevance of defensive information. Conceded xG, conceded goals, xBG, and conceded EPV contribute to the models, showing that predicting a team’s future scoring output depends not only on its own attack but also on the defensive quality of the opponent. This is especially relevant because defensive analytics remains less developed and less represented in the literature than attacking metrics.
When compared with previous match prediction studies, the authors report that their best pre-match and post-match models perform well relative to approaches from Berrar et al. [18], Hubáček et al. [13], and the 2023 soccer prediction challenge [12]. However, they caution that direct comparisons are difficult because studies differ in leagues, seasons, input data, prediction horizons, and modelling objectives. Their model is specifically designed to predict the next Bundesliga matchday with continuous updates from recent data, rather than forecasting across many leagues or entire seasons.
4.1 Practical application
The authors argue that pre-match and post-match models serve different roles, and their greatest value comes from combining them.
In pre-match analysis, the EPV-based model can estimate win, draw, and loss probabilities before the game. This can support opponent analysis, expectation-setting, match planning, and trend monitoring. It can also be extended to predict future league standings, such as winter break or final table positions, based on current performance trends.
In post-match analysis, the xG-based model provides an objective evaluation of the performance produced in the match. It helps analysts compare the actual result with the quality of chances created and conceded. This supports a common club-level need: separating performance from outcome.
The most interesting use case comes from comparing pre-match and post-match probabilities. Figure 3 shows an example where the pre-match EPV prediction and post-match xG prediction are placed side by side. The authors show how a team’s win probability can shift after the match based on the performance produced, even if the final result does not fully reflect that performance. In the example, the home team increased its win probability by 9% and reduced its loss probability by 16% after outperforming the opponent in xG, but the match still ended 2:2.
The authors also note that SHAP can make these models more useful in practice. Instead of only saying that a team is predicted to score a certain number of goals, SHAP can explain which features drove that prediction. For analysts and coaches, this interpretability matters because model outputs are easier to trust when they can be connected to football concepts such as home advantage, opponent strength, recent attacking form, or defensive vulnerability.
4.2 Limitations and future research
The first limitation is that the models do not include individual player information. The authors do not account for starting line-ups, injuries, suspensions, player-specific xG or EPV contributions, or tactical changes linked to personnel. This is a meaningful limitation because team strength can change substantially depending on which players are available and how they are used.
The second limitation concerns the use of independent Poisson distributions to model goal outcomes. Although this approach is common and robust, it can underestimate draws. This is particularly relevant because 1:1 was the most frequent scoreline in the sample. Future research could test alternative distributions or more advanced models, including Bayesian hierarchical approaches or temporal neural networks that account for uncertainty, team-specific effects, and evolving team strength.
The third limitation is the narrow contextual feature set. The models include match venue, but not injuries, coaching changes, fixture congestion, season phase, or other situational factors. Including such variables could improve predictive performance and make the models more sensitive to real-world football context.
The fourth limitation is external validity. The dataset covers only three Bundesliga seasons, so it is unclear whether the same relationships would hold across longer time periods, other leagues, cup competitions, or international tournaments. Future research should test whether EPV remains stronger for pre-match prediction and xG remains stronger for post-match prediction across different competitions and data environments.
Finally, the authors note a general interpretability issue with machine learning models. Unlike traditional statistical models, machine learning approaches do not have universally accepted thresholds such as p-values or effect sizes. Feature importance values can help, but interpretation becomes harder when several features have similar importance.
5 Conclusion
The paper concludes that xG and EPV are both valuable, but for different analytical tasks. EPV is more useful in pre-match prediction because it captures broader possession-level information and therefore appears better suited to forecasting future team performance. xG is more useful in post-match prediction because it directly summarizes the quality of the shots that actually occurred, making it a strong measure of match performance after the game has been played.
The central applied message is that analysts should not treat KPIs as universally superior or inferior. Their value depends on the use case. EPV helps describe and predict the process leading into future matches. xG helps evaluate the chances created in a match that has already happened. Used together, pre-match EPV and post-match xG can help clubs compare expectation, performance, and result, giving coaching staff a more objective way to assess whether a team’s outcome was supported by the underlying performance.
Learn More
Here you have a direct link to get the free xG Mini-Guide, suited for those who want to fully understand the model in 5 minutes.
References
Forcher, L., Forcher, L., Woll, A., & Altmann, S. (2025). AI in Bundesliga match analysis—expected possession value (EPV) vs. expected goals (xG) to predict match outcomes in soccer. Frontiers in Sports and Active Living, 7, 1713852.
https://doi.org/10.3389/fspor.2025.1713852